Neural networks that borrow strategies from biology are making profound leaps in their abilities. Is ignoring a goal the best way to make truly intelligent machines? A new paper proposes to find out by training neural networks to look ahead of their next move, allowing them to build strategies from a previous move.
The idea is that neural networks should be trained using the best move of a previous round — or, in the jargon of the study, “the last learned move.” This way, when a new game is introduced, neural networks can learn to anticipate the outcome by looking backward in time.
The approach seems to be catching on, as researchers at MIT, the University of California, Berkeley, and Harvard have been using this method for training neural networks.
Read the full story at …
The Offworld Report – the best content offworld SFcrowsnest.